Fig. 3: Learning compound-CRISPR perturbation representations.
From: In silico biological discovery with large perturbation models
a, The latent space of compound and CRISPR knockouts (reduced to two-dimensions via t-SNE) reflects known groupings of compound and genetic perturbations that target the same molecular mechanisms in bulk LINCS L1000 data from ref. 7. Genes targeted by corresponding CRISPR and compound inhibitors are color-coded in matching colors. b, Root mean squared error (RMSE) distances of known HMGCR inhibitors (statins) to the corresponding CRISPR-HMGCR perturbation in the embedding space of the LPM. Two bottom outliers are additionally annotated in a: benfluorex (withdrawn for cardiovascular side effects42) and pravastatin (shown to have low correlation to other statins78 and additional anti-inflammatory effects44,45,46). c, The RMSE-based distance between perturbation embeddings for CRISPR perturbations was used to measure the recall of known inhibitors of the respective genetic target, for different numbers of nearest neighbors. We compared LPM embeddings with those derived from post-perturbation L1000 transcriptome profiles. Bars represent the 95% confidence intervals across genetic targets (N = 89).
